rollback 2 days

app.py

@@ -1,44 +1,53 @@
- app.py – virkar 100% á ZeroGPU (og venjulegum GPU líka)
 import os
+os.environ["OMP_NUM_THREADS"] = "1"                     # Fixes libgomp + asyncio FD errors
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"  # Prevents VRAM crash
+
 import gradio as gr
 import spaces
 from transformers import pipeline
 
 MODEL_NAME = "palli23/whisper-small-sam_spjall"
 
-# Látum mótið hlaðast EINUSINNI þegar appið ræsist (ekki í hverju kall)
-print("Hleð Whisper módelinu einu sinni...")
-pipe = pipeline(
-    "automatic-speech-recognition",
-    model=MODEL_NAME,
-    torch_dtype="auto",
-    device_map="auto",           # þetta er lykillinn fyrir ZeroGPU
-    token=os.getenv("HF_TOKEN")
-)
-print("Módel tilbúið!")
-
-@spaces.GPU(duration=120)   # 2 mínútur nægja fyrir 3–4 mín hljóð
+@spaces.GPU(duration=180)  # Extra buffer for cold start
 def transcribe_3min(audio_path):
     if not audio_path:
-        return "Hladdu upp hljóðskrá fyrst"
+        return "Hladdu upp hljóðskrá"
+    
+    pipe = pipeline(
+        "automatic-speech-recognition",
+        model=MODEL_NAME,
+        device=0,
+        torch_dtype="float16",        # Keeps speed + enables real batching
+        token=os.getenv("HF_TOKEN")   # Optional – remove if no private model
+    )
     
     result = pipe(
         audio_path,
         chunk_length_s=30,
-        batch_size=8,
+        stride_length_s=(6, 0),
         return_timestamps=False,
-        generate_kwargs={"language": "is", "task": "transcribe"}
+        batch_size=8
     )
+    
     return result["text"]
 
+
 with gr.Blocks(title="Íslenskt ASR – 3 mín") as demo:
-    gr.Markdown("# Íslenskt Whisper – 3 mín hljóð á ZeroGPU")
-    gr.Markdown("Hladdu upp allt að 3–4 mínútna hljóðskrá (mp3/wav). Virkar strax!")
+    gr.Markdown("# Íslenskt ASR – 3 mínútur")
+    gr.Markdown("**Whisper · Very low WER · 0.5–5 mín hljóð á ZeroGPU**")
+    
+    audio = gr.Audio(type="filepath", label="Hladdu upp .mp3 / .wav (allt að 5 mín)")
+    btn = gr.Button("Transcribe", variant="primary", size="lg")
+    out = gr.Textbox(lines=30, label="Útskrift")
     
-    audio_in = gr.Audio(type="filepath", label="Hljóðskrá (max ~4 mín)")
-    btn = gr.Button("Transcribe", variant="primary")
-    text_out = gr.Textbox(lines=25, label="Útskrift")
+    btn.click(transcribe_3min, inputs=audio, outputs=out)
 
-    btn.click(transcribe_3min, inputs=audio_in, outputs=text_out)
 
-demo.launch()
+demo.launch(
+    auth=("beta", "beta2025"),
+    ssr_mode=False,       # Critical – stops infinite loading loop
+    show_error=True,      # Shows real errors instead of blank screen
+    server_name="0.0.0.0",
+    server_port=7860,
+    quiet=False
+)